Claude Opus 5 Beats Legend 5 in Most Benchmarks—At Half the Price

Claude Opus 5 is out today. It’s cheaper for businesses to run than Anthropic’s flagship model, the Claude Fable 5, which the company listed as an everyday product for paying users. In addition, Opus 5 also outperforms benchmarks.

To understand where Opus 5 fits in: The Anthropic line consists of four parts. Haiku is fast and cheap. The sonnet is medium. Opus is the heaviest horse. Above that is the Mythos group – the Anthropic part that was introduced in the spring – which includes Claude Fable 5 to the public, and Claude Mythos 5, a version with limited restrictions reserved through Project Glasswing for cyber security researchers and critical infrastructure workers.

Fable 5 has been struggling as a subscriber. Launched on June 9, it was he came to the world after three days after the US government issued an emergency export ban citing the risk of prison, and it came back on 30 June-immediately switching to a single model, no longer included in standard plans. Opus 5 is now filling the space Fable 5 couldn’t hold.

Beloved, a construction platform with millions of users, ran Opus 5 in its internal evaluations and found that the gains go beyond that: “It is not better in our most demanding tasks, up to 22% more than Opus 4.7, it is more stable, with less difference,” Fabian Hedin shared in the words of Anthropic.

Measurements

It may sound strange, but Opus beats Fable in almost everything that would be important to a daily user while not being Mythos-class like Fable.

On Frontier-Bench v0.1—a benchmark that measures whether AI code assistants can complete real end-to-end software engineering tasks, expressed as the number of tasks passed—Opus 5 scored 43.3%. Legend 5 came in at 33.7%. OpenAI’s GPT-5.6 Sol, Anthropic’s biggest commercial rival, scored 34.4%.

The biggest edge is on the ARC-AGI-3, a real-world problem-solving test designed around new puzzles that the model could not have memorized from the training course, scored as the number of puzzles solved. Opus 5 hit 30.2%; GPT-5.6 Sol gained 7.8%; and Fable 5 was never tested. On the GDPval-AA v2—an indicator of the performance of information generated through Elo ratings, a chess system used to measure real-world performance—Opus 5 reached 1,861 against Fable 5’s 1,747 and GPT-5.6 Sol’s 1,736.

Zapier tested Opus 5 on AutomationBench, a test that shows whether a brand can handle a complete workflow from start to finish without human intervention. Their verdict: the model “took a raw account health manual and ran an end-to-end churn process: identifying the accounts at risk, alerting the right owner, and summarizing the retention ops. The previous models did not pass; Opus 5 hit 100%.”

Anthropic is also positioning Opus 5 as a research upgrade. Ultima Genomics, a DNA sequencing company, said the model “behaves more scientifically than any model we’ve ever run. It reaches the right statistical tests to rule out confounders, checks results with independent methods, and stays on track through multivariate analysis.”

That said, the two areas – law and health – are the only ones where Fable 5 slightly excels.

The release comes a week after Moonshot AI, a Beijing-based startup backed by Alibaba, unveiled the Kimi K3 – a 2.8-trillion-parameter open-weight model (meaning anyone can download a small piece of code to run it on their own) which Moonshot describes as the world’s largest open-source AI system. Independent benchmarks place the Kimi K3 in third overall, behind the Fable 5 and GPT-5.6 Sol, beating the two in some areas.

Opus 5 is now available through the API at $5 per million input tokens and $25 per million output. (Signs are an important part of the knowledge that an AI model can process for input and output). A Fast option running at about 2.5 times the default speed is also available, at twice the price – $10 per million input tokens and $50 per million output.

This release may put an end to concerns over the lack of Fable 5 in the community. Opus is also available by subscription to everyone.